Case study · Rust · v0.12.0 beta · AGPL-3.0
PeerBeam
Open the app, see nearby devices, click, send. No IP addresses, no pairing codes, no accounts, no cloud — across LAN, mDNS and Tailscale at once.
The problem
LAN transfer tools ask the user to do the network's job: type an address, read out a code, create an account. And they assume both devices are on the same LAN — the moment one is on a tailnet, behind a VPN, or is a headless server reached over SSH, the tool has nothing to offer.
PeerBeam removes the configuration rather than documenting it. Discovery runs over LAN broadcast, mDNS and the Tailscale LocalAPI simultaneously, and the results merge into one device list. A device is listed by the name it chose, not by where it happens to be reachable.
Two frontends, one engine
The core is a Rust workspace following Clean Architecture: dependencies point inward toward peerbeam-domain, which defines the ports — traits — that every provider implements. Discovery, transport, storage, crypto and trust are all adapters behind those traits, so adding Bluetooth discovery or a WebRTC transport is a new crate rather than a change to the engine.
A Flutter app and a scriptable CLI sit over that one core, the app reaching it through a stable C-ABI bridge. The CLI is not a debugging afterthought — it is the frontend that works over SSH on a machine with no display.
The CLI
One discover returning a tailnet peer and a LAN peer in a single list, then a send to one of them by name:
A peerbeam discover command lists two devices — one reachable remotely over Tailscale, one on the LAN — followed by a send command that pins the new peer, prints a pairing code, and reports the file sent.
The pairing code is not a thing to type. It is a fingerprint shown once on both devices so they can confirm they pinned the same key.
Security that is on by default
Key agreement is X25519 ECDH into directional session keys; every frame is sealed with AES-256-GCM and carries a replay counter. A peer's fingerprint is pinned on first contact, and a changed fingerprint later is treated as what it is — something to stop for, not a prompt to click through.
Approval and permission are kept as separate questions. A device can be trusted for files and still refused the clipboard, and clipboard sync is opt-in on both ends, because everything copied is sent — passwords included, which the toggle says plainly rather than guessing what is sensitive.
Numbers
1.6 ms
CLI startup
median, peerbeam --version
~13 MB
RAM per transfer
flat at any file size
2,633
tests
1,836 Rust · 797 Flutter
~5 ms
discovery overhead
discover --timeout 2
Transfer RAM staying flat at any file size is the one worth pausing on: it is the evidence that the pipeline really streams and never holds a whole file in memory.
The profiling also produced a result worth keeping. AES-256-GCM runs at roughly 900 MiB/s sealing and 1,010 MiB/s opening on a single core, while an unencrypted in-process loopback transfer manages about 490 MiB/s. The encryption is faster than the pipeline around it — so the bottleneck is hashing, not crypto, and time spent optimising the cipher would have bought nothing.
What is not done
v0.12.0 is beta, and the gap to a stable 1.0 is verification rather than features. Linux, Android and the CLI are build- and run-verified, and one real cross-device transfer is confirmed byte-exact between Android and Linux. Windows and macOS are built and analysed by CI but not exercised by hand, so notification delivery and file staging on those two platforms are listed as unverified rather than assumed to work.
The project's own release gate says the same thing in stronger terms: shipping “stable” while whole platforms are untested would inflate quality against evidence that does not exist yet.